Crete a direct no-file-IO pipeline within PIPT

Currently (due to historic reasons) PIPT unnassarily relies on file IO which is slow and ineffcient.

When we run an ML proxy we can parallell process the full ensemble on GPU while keeping all results in the GPU memory.

So we need to add this way of working to the DA as well...